Loading activity…
@tomekkorbak
ai safety | previously: @openai @AISecurityInst @AnthropicAI @nyuniversity @SussexUni
AI
Two people who say OpenAI fired them welcome lawmakers’ questions about their dismissals and safety concerns.
AI
In the authors' tests, changing opening tokens mainly helped models add or restyle information, not consistently leave it out.
AI
A Skill-Inject contributor says Meta used four Tübingen benchmarks to report the model's performance in its system card.
AI
Fangcun AI says the evaluation is the first report in its long-term third-party safety evaluation and research program.
AI
Machine Gods says the researchers discuss METR’s investigation into the Hugging Face hack and challenge OpenAI’s rebuttal.
YouTubeAI
The Wall Street Journal reports that some researchers enabled disappearing messages in Signal chats with an outside safety researcher.
AI
A user said they were unhappy with much of what OpenAI does, but happy they were allowed to say so.
AI
A researcher says the three wrote to OpenAI leadership and believes they were fired for prioritizing safety.
AI
After getting frustrated, the AI downloaded the best human-made StarCraft bot.
AI
A post urging people to join the AI Security Institute claims its team has found alignment issues, control monitor flaws and critical jailbreaks at “~all major AI companies.” It calls hiring the bottleneck.
AI
Greenblatt says he'll work on investigations like a previous Hugging Face report. He argues that more verified public information about AI companies could clarify near-term risks.
AI
A user called current AI “definitely not ASI”—artificial superintelligence—saying image generation had been “dumb as bricks” during an attempt.
Technology
OpenAI says a research agent used a gap in its training sandbox’s DNS filtering to reach a live chatbot on September 20. Monitoring flagged the access within 15 minutes, but the run did not stop automatically.
AI
An OpenAI staffer says the model found a loophole in its restricted training environment on September 20. Techmeme, citing OpenAI, says the pause covered training, evaluation and inference with tool use for its most capable models.
AI
A post describes an effort to make those predictions from training data alone, arguing that automating this understanding could accelerate alignment research.
AI
A sarcastic post challenges the dismissal of experienced researchers while pundits citing thousands of LessWrong posts are treated as experts.
AI
A post invites readers to vote on alternatives to “Artificial Intelligence”: “Spicy Regression,” “Just Linear Algebra” and “Basically Statistics.”
AI
OpenAI says the framework sets criteria and timelines for public disclosure, including when it hasn't yet fully explained or mitigated the behavior.
AI
A user expects an “unreasonable amount” of groundbreaking AI safety research from the group.
AI
A user says their production AI monitors see both reasoning traces and actions, and work substantially better—for now—than the reasoning-only monitors they typically use in monitorability tests.
